We pursue research in machine learning and optimization. To this end, we develop theories and algorithms using computational and mathematical tools. Our ultimate goal is to provide robust and provable solutions to challenging problems in artificial intelligence, particularly those in large-scale settings. We are passionate about translating our findings into practical applications that can benefit society. For further details on our research directions and ongoing projects, please refer to the Research.
Recent News
Oct 2026
🎓 Our preprint on decentralized LLM training, SeedFlood, has received a major update, with substantial improvements to its overall presentation, analysis, and experimental evaluation.
🎓 Our preprint on decentralized LLM training, SeedFlood, has received a major update, with substantial improvements to its overall presentation, analysis, and experimental evaluation.
Oct 2026
🧑💻 Minkyoung has started a research internship at Nota AI.
🧑💻 Minkyoung has started a research internship at Nota AI.
Oct 2026
🎓 Our paper on hardware-native sparse-quantization for trillion-scale MoEs is out on arXiv. In this work, we introduce MoESQ, which jointly learns semi-structured sparsity and low-precision quantization for MoE experts and executes them on Sparse Tensor Cores with a custom grouped sparse GEMM kernel. The code and models are publicly available.
🎓 Our paper on hardware-native sparse-quantization for trillion-scale MoEs is out on arXiv. In this work, we introduce MoESQ, which jointly learns semi-structured sparsity and low-precision quantization for MoE experts and executes them on Sparse Tensor Cores with a custom grouped sparse GEMM kernel. The code and models are publicly available.
Sep 2026
🎓 Our paper on decentralized training has been accepted to NeurIPS 2026 Workshop on Collaborative, Open, and Decentralized Training of Foundation Models, selected as an oral presentation. In this work, we introduce a novel decentralized LLM fine-tuning framework that achieves extreme communiation efficiency and global consensus.
🎓 Our paper on decentralized training has been accepted to NeurIPS 2026 Workshop on Collaborative, Open, and Decentralized Training of Foundation Models, selected as an oral presentation. In this work, we introduce a novel decentralized LLM fine-tuning framework that achieves extreme communiation efficiency and global consensus.
Sep 2026
🎓 Our paper on asynchronous pipeline parallelism has been accepted to NeurIPS 2026. In this work, we introduce FRESCO, a new principle for consistency control in asynchronous pipelines, which consistently outperforms standard baselines while being extremely efficient and theoretically sound.
🎓 Our paper on asynchronous pipeline parallelism has been accepted to NeurIPS 2026. In this work, we introduce FRESCO, a new principle for consistency control in asynchronous pipelines, which consistently outperforms standard baselines while being extremely efficient and theoretically sound.
If you want to see our earlier news, check out here!
